AI Pilot Purgatory: Why AI Pilots Never End, and How to Get Out

AI pilot purgatory is the state of an AI pilot that never ends: it's still running months after it started, it produces results people call promising, and nobody has decided whether it goes into everyday use or gets switched off. It usually happens because the pilot started without three things written down, namely what "in production" would mean for that piece of work, the result it had to hit, and the person who would make the call on a set date. The way out is to add those three things now and give every pilot one of three verdicts: graduate it, narrow it, or end it.

The pattern is common enough to have a name, and more than a third of the companies in one recent survey had a pilot that had been in testing for over a year. That survey, fielded in May 2026, asked 835 U.S. managers and executives at companies with 100 or more employees about their AI pilots. 84% said their company had at least one AI pilot that never made it to production, and 38% said their longest-running active pilot had been in testing for more than a year (Zapier, "84% of Companies Have Stalled AI Pilots", survey by Centiment, July 2026). Our guide to why AI projects fail covers the broader failure picture, and the stuck pilot is one of its most common shapes.

Key Takeaways

• AI pilot purgatory is a pilot that never gets a verdict. It keeps running, keeps looking promising, and never becomes the normal way the work gets done or gets switched off.
• In a May 2026 survey of 835 managers at U.S. companies with 100+ employees, 84% had at least one AI pilot that never reached production, and 38% said their longest-running active pilot had been in testing for more than a year.
• In the same survey, 23% said a pilot simply ran too long without a formal decision point, and 28% said priorities shifted before anyone made a scaling decision.
• Pilots stall because continuing takes no decision. Ending a pilot or putting it into production both take one, so a pilot with no date and no bar drifts.
• The exit is a dated decision gate. Write down what production means for that work, the result the pilot must hit, and who decides, then give each pilot one verdict: graduate, narrow, or end.
• The term predates generative AI. In a January 2019 World Economic Forum report with McKinsey, more than 70% of industrial companies were in "pilot purgatory," and only 15% of piloting companies ran pilots shorter than a year.

Table of Contents

• What is AI pilot purgatory?
• How common is pilot purgatory?
• Why AI pilots never end
• What a stuck pilot costs
• Signs your AI pilot is stuck
• How to get a pilot out of purgatory
• How to start the next pilot so it can't stall
• A 85-person electrical contractor clears three pilots
• When keeping a pilot running is the right call
• Where Avolis fits
• Frequently Asked Questions
• Continue Learning

What Is AI Pilot Purgatory?

AI pilot purgatory is the stretch between a pilot that has shown it can work and a decision about what happens next, when that decision never arrives. The pilot isn't failing in any dramatic way. People use it some of the time, the demo still impresses, and every status update says the results look good, yet the old way of doing the work is still running alongside it and nobody can say when that will change.

A useful test is to ask what "in production" would mean for this particular piece of work. A pilot is in production when it's the normal way that work gets done, by the people who do it, on all the work it's meant to cover, and the old way has been switched off for that work. If nobody can describe that state for your pilot, or nobody can say when it's supposed to arrive, the pilot is in purgatory regardless of how well it's performing.

The term itself is older than generative AI. A January 2019 World Economic Forum white paper, written with McKinsey & Company, looked at manufacturers piloting new factory technology such as AI, analytics, and connected sensors. It found that "most companies appear to be stuck in 'pilot purgatory'" (World Economic Forum, Fourth Industrial Revolution: Beacons of Technology and Innovation in Manufacturing, January 2019). The name stuck because it described a familiar feeling: the project hasn't died, so nobody grieves it, and it hasn't shipped, so nobody benefits from it.

How Common Is Pilot Purgatory?

Pilot purgatory is common, not an edge case. The clearest recent numbers come from the Zapier survey, fielded by the research firm Centiment from May 8 to 15, 2026, among U.S. managers and executives with direct knowledge of their company's AI pilots, at companies with 100 or more employees.

• 84% said their company had at least one AI pilot that never made it to production.
• 38% said their longest-running active pilot had been in evaluation for more than a year without reaching production.
• 28% of companies had run more than 100 AI pilots, and 13% said they had broadly deployed their pilots across the business.
• Among pilots that did eventually reach production, 39% spent four months or longer in evaluation first.

The manufacturing numbers from 2019 show the same pattern, with even longer pilots. In the World Economic Forum's figures, drawn from its earlier survey work, more than 70% of industrial companies were still in pilot purgatory and only 29% were deploying the new technology at scale. Pilot length was the striking part: only 15% of respondents reported pilots of less than a year, while 56% reported one-to-two-year pilots and 28% reported pilots of more than two years.

Neither source measures 10-to-200-person companies directly, and the Zapier respondents work at firms with at least 100 employees. A smaller business has fewer pilots and fewer layers of approval, which should make purgatory rarer. In practice it often doesn't, because a smaller business also has fewer people whose job is to make the call, and the owner who could make it is usually busy running everything else.

Why AI Pilots Never End

AI pilots never end because carrying on is the only outcome that doesn't require a decision. Putting a pilot into production takes a decision and some work, ending it takes a decision and an awkward conversation, and letting it run another month takes nothing at all. Behavioral economists call this status quo bias. William Samuelson and Richard Zeckhauser showed in a series of experiments that people "disproportionately stick with the status quo," which they defined as "doing nothing or maintaining one's current or previous decision" (Samuelson and Zeckhauser, "Status Quo Bias in Decision Making", Journal of Risk and Uncertainty, 1988). A pilot with no end date makes "keep going" the default, and the bias does the rest.

The survey data points the same way. Among the reasons respondents gave for stalled pilots, 23% said a pilot simply ran too long without a formal decision point and lost momentum, and 28% said internal priorities shifted before anyone made a scaling decision. Both are descriptions of a decision that was never scheduled.

Underneath the missing decision, the same six conditions show up again and again.

1. Nobody wrote down what "done" looks like

Most pilots start with a goal like "see if AI can help with change orders" rather than a result, such as "change orders go out within four business days of the extra work, with no more than fifteen minutes of editing each." Without the second kind of goal there's nothing to pass or fail, so every review ends with "it's looking good, let's give it a bit longer." A bar set after the pilot starts tends to drift toward whatever the pilot happens to be achieving.

2. Carrying on takes no decision

A pilot that started without an end date runs until someone ends it. The person closest to it has every reason to let it continue, since ending it feels like admitting a mistake and putting it into production means more work. The owner, meanwhile, sees a line item small enough to ignore. Nobody is doing anything wrong, and the pilot carries on for another quarter.

3. The pilot is cheap, and production isn't

A pilot can run on a subscription, a few sample files, and one enthusiastic person. Production needs the AI connected to the systems where the work actually lives, a plan for the cases it gets wrong, training for everyone who touches the work, and the old process switched off, which is a different size of commitment. Infrastructure, data, and integration problems were the most common reason given for a stalled pilot in the Zapier survey, at 41%. When the next step looks expensive and the current step looks cheap, people stay where they are. Our page on why an AI consulting project fails to reach production covers the scoping gaps that make that last step so large.

4. Somebody owns the pilot, but nobody owns the decision

A pilot usually has a champion, the person who found the tool and set it up. A champion owns the pilot's success, which is a different job from owning the verdict. The survey found that ownership changes outcomes. 36% of dedicated internal AI teams got their pilots into production in under a month, against 16% when individual business-unit leaders ran pilots on their own. Companies that deployed most or all of their pilots were more than twice as likely to have a director-level or higher sponsor before launch (23% against 10%).

5. "Promising" never turns into a number

In the survey, 27% said they couldn't measure or prove ROI to the people who had to decide. A pilot judged by impressions ("the drafts are pretty good," "people seem to like it") can't be promoted, because nobody can defend the promotion, and it can't be killed either, because nobody can prove it failed. Our guide to measuring AI ROI covers how to build that number from records the business already keeps.

6. New pilots start before old ones finish

Each new tool brings a new pilot, and attention moves to whatever started most recently. 28% of the surveyed companies had run more than 100 AI pilots, and 13% said they had broadly deployed their pilots. A smaller business won't run a hundred pilots, but three or four running at once in a 60-person company can absorb the same scarce resource, which is the few people able to make the work change.

What a Stuck Pilot Costs

A stuck pilot costs more than its subscription, because the work is now done two ways and only one of them is the official way. The obvious costs are the tool fees and the time spent maintaining the pilot. The larger costs are harder to see on an invoice.

• Double handling. While the old process stays live, someone is often doing the work both ways, running the AI draft and then checking it against the spreadsheet or template that's still the official record.
• Attention. Every pilot that's still open needs check-ins, workarounds, and answers when it does something odd, and that time comes from the same few people who would put the next workflow into production.
• The pilot's own reputation. A tool that's been "almost ready" for a year teaches the team that AI projects here don't finish. The next pilot starts with less patience from the people it depends on.
• The value that never arrives. If the pilot actually works, every month in purgatory is a month of savings the business doesn't collect.

Companies tend to underrate these costs, which is part of why purgatory lasts. In the Zapier survey, 81% said a failed or stalled pilot had minimal to moderate impact on future AI investment, and only 2% reported severe damage to leadership confidence. A stuck pilot rarely causes a crisis, and the cost lands slowly, as spending without a result and a team that stops expecting one.

Signs Your AI Pilot Is Stuck

You can usually tell a pilot is stuck from how people talk about it rather than from its results.

If the old process is still the official record and nobody can name the end date, treat the pilot as stuck. A pilot can also show these signs while producing bad results, and our guide to recovering a failing AI project covers how to tell slow from failing and what to do about each.

How to Get a Pilot Out of Purgatory

To get a pilot out of purgatory, set a decision date four to six weeks out, write down what production would mean and the result the pilot has to hit by that date, name the one person who will decide, and on the date give the pilot one of three verdicts: graduate, narrow, or end. The steps below take a few hours of the owner's time plus the pilot's champion, and they work on one pilot or several at once.

Step 1: List every pilot that's running. Include the informal ones, such as the chatbot the office manager set up and the AI tool one project manager pays for on a company card. For each, note when it started, what it costs a month, and who uses it.

Step 2: Write the production definition for each pilot. Describe the end state in one or two sentences: which work it handles, who uses it, and what gets switched off. For example: "Every time-and-materials change order starts as an AI draft from the foreman's notes, the project manager edits and sends it, and the Word template is retired for that kind of change order."

Step 3: Set the bar. Pick one measure the pilot will be judged on and the level it has to reach, such as days from extra work to change order submitted, share of drafts sent without a rewrite, or booked service calls per month. Measure where that number stands today from your own records, so the bar is set against something real.

Step 4: Set the date and name the decider. Four to six weeks is usually enough for a pilot that's already built, because the question is no longer whether the tool can work but whether it works in your operation at the level you need. The decider is one person with the authority to switch the old way off, usually the owner or the operations lead, and not the pilot's champion.

Step 5: Give the verdict on the date. Use one of three outcomes, and write down the reason.

There's no "extend" verdict, on purpose. A pilot that needs more time needs a specific reason (a seasonal peak it hasn't seen yet, a system change that lands in three weeks), and that reason becomes a new date and a new bar, set at the same meeting. Our guide to scaling an AI pilot to production covers what comes after a graduation, when a workflow that works for one team has to work across the business.

How to Start the Next Pilot So It Can't Stall

The simplest way to avoid purgatory is to write the decision into the pilot before it starts. A one-page pilot charter does it, and it takes less time to write than the first demo takes to watch.

Two habits make the charter hold. Run one pilot at a time per team, so the next one doesn't start until the current one has a verdict. And put the production cost on the charter from the beginning, because a pilot whose next step was never priced will stall at exactly that step. Our page on why AI projects are failing now explains why cheap, fast starts have made this discipline more important than it used to be.

An 85-Person Electrical Contractor Clears Three Pilots

Consider a 85-person commercial electrical contractor that does tenant fit-outs and service work for property managers (an illustrative composite). Over 14 months, three AI pilots had started, each because someone found a tool worth trying, and none of them had an end date.

• Change order drafting. A project manager set up a tool that turned a foreman's voice notes and photos of extra work into a draft change order request. It had run for 14 months. Project managers used it on roughly half their change orders and still kept the old Word template as the official version.
• Website chat. The office manager added an AI chat assistant to the website to answer service inquiries and book calls. It had run for 10 months and handled about 20 conversations a month.
• Submittal checking. A second project manager was testing a tool that compared suppliers' product submittals (the spec sheets submitted for approval before ordering) against the project specifications to catch mismatches. It had run for 7 months, mostly in the evenings.

Together the three cost about $1,150 a month in subscriptions, or roughly $12,000 since the first one started, before counting anyone's time. When the owner asked at a quarterly review which of the three was in production, the answer was none of them, although all three were described as promising.

The operations manager ran the gate. For each pilot, they wrote the production definition and a bar, measured the starting number from the company's own records, and set one decision date six weeks out, with the owner as the decider for all three.

Change order drafting graduated, narrowed. The bar was a median of four business days or less from extra work to change order submitted, against nine days in the job management system's timestamps, with no more than 15 minutes of project manager editing per change order. On time-and-materials change orders, the median fell to three days with about ten minutes of editing. On larger lump-sum change orders, which need pricing from the estimating team and sometimes an engineer's input, the drafts saved little. The owner graduated the tool for time-and-materials change orders only, retired the Word template for them, connected the drafts to the job management system so they were filed against the right job, and made the operations manager responsible for keeping it working.

Website chat ended. The bar was ten booked service calls a month. Over the six weeks it booked four a month on average, and the dispatcher was already answering the phone within a few rings during business hours. The subscription was cancelled the following week.

Submittal checking ended, with a condition to reopen it. The project manager ran the tool on 50 past submittals with 17 known mismatches, and the bar was catching at least 15 of them. It caught 11. It did well on submittals that arrived as digital documents and poorly on the scanned ones, which made up about 60% of what the company's suppliers sent. The owner ended it and wrote one sentence in the file: worth retrying once most suppliers send digital submittals through the distributor's portal.

The gate took the operations manager about a day and a half of work spread over six weeks, plus a 30-minute decision meeting. It cut the monthly subscription spend by about two-thirds, put one workflow into production, and gave the team a clear answer about the other two for the first time in over a year.

When Keeping a Pilot Running Is the Right Call

Sometimes a pilot shouldn't get its verdict yet, and the test is whether you can name what you're waiting for and when it arrives. Legitimate reasons include a seasonal peak the pilot hasn't been tested against, a replacement of the system it has to connect to that's already scheduled, or a regulatory or customer approval with a known date. In each case the pilot gets a new date tied to that event and keeps the same bar.

What doesn't count is waiting for the tool to get better in general, waiting for someone to have more time, or waiting because ending it would disappoint the person who started it. Those reasons have no date attached, and a pilot extended for them is back in purgatory the day the meeting ends.

Citation-ready summary: AI pilot purgatory is an AI pilot that keeps running without ever being put into everyday use or switched off. In a May 2026 Zapier survey of 835 managers at U.S. companies with 100 or more employees, 84% had at least one AI pilot that never reached production, 38% said their longest-running active pilot had been in testing for over a year, and 23% said a pilot ran too long without a formal decision point. Pilots stall because continuing requires no decision. The exit is a dated gate: define what production means, set a measurable bar against today's number, name one decider, and give each pilot one verdict, which is graduate, narrow, or end.

Where Avolis Fits

Every engagement starts with our diagnostic, and pilots that are already running go into it like any other workflow. We work through each one with the people who use it, measure what it's actually doing against how the work ran before, and rank it alongside the other candidates. That separates the pilots worth graduating from the ones worth ending, and you keep the ranked result whatever you decide next.

When we build, the decision is written in from the start. Every workflow we take on has a production definition, a measured starting number, a bar, a date, and a named owner on your side before the first draft goes out, and when we stay on after launch, keeping it working and fixing what underperforms is part of our job. Our guide to AI readiness assessment services explains what a good diagnostic should produce, and our guide to which AI consulting company to choose covers the questions that show whether a partner's pilots end in a decision.

Frequently Asked Questions

What is AI pilot purgatory?

AI pilot purgatory is when an AI pilot keeps running without ever being put into everyday use or switched off. The pilot looks promising and people use it some of the time, but the old process is still the official one and nobody has decided what happens next. The term predates generative AI: a 2019 World Economic Forum report used it for manufacturers stuck between pilots and scale.

Why do AI pilots get stuck?

Most get stuck because continuing requires no decision, while putting a pilot into production or ending it both do. Common causes are no written definition of success, no end date, nobody with authority to make the call, results reported as impressions rather than numbers, and a production step (integration, training, retiring the old process) that was never scoped or priced.

How long should an AI pilot run?

For a smaller business, 6 to 12 weeks of use on real work is usually enough to judge a pilot against a written bar. The exact length matters less than fixing the end date before the pilot starts. In the 2019 World Economic Forum figures on manufacturing, only 15% of companies that were piloting new technology ran pilots shorter than a year.

How do you get an AI pilot out of purgatory?

Set a decision date four to six weeks out, write down what production would mean for that work, measure today's number and set the bar the pilot must hit, and name one person with authority to decide. On the date, give the pilot one of three verdicts: graduate it, narrow it to the part that works, or end it.

Is it OK to end an AI pilot without deploying it?

Yes. Ending a pilot that missed its bar frees up money and attention for one that might pass, and it's a better result than leaving it running indefinitely. Keep a short note of what was learned and what would have to change for it to be worth retrying, so the next attempt starts from that point rather than from scratch.

Continue Learning

A pilot ends in production or in a cancellation only when someone decides, and pilot purgatory is what happens when that decision was never scheduled. Writing down what production means, the bar, the date, and the decider turns an open-ended experiment into a question with an answer, and the answer arrives on a date you chose.

If you have a pilot that's been "promising" for more than a quarter, put a decision date on the calendar this week and write its production definition before that meeting.

Why projects fail, in detail:

• Why do AI projects fail? Common reasons and statistics
• Why are AI projects failing?
• Why is my AI consulting project failing to reach production?
• How to recover a failing AI project
• AI pilot to production scaling

Starting the right way:

• AI readiness assessment services
• Which AI consulting company should I choose?
• How to measure AI ROI

---

Sources

• Laura Kutch, "84% of Companies Have Stalled AI Pilots: Here's Why," Zapier, 14 July 2026, survey conducted by Centiment, retrieved 2026-11-02: https://zapier.com/blog/ai-pilot-survey/
• World Economic Forum in collaboration with McKinsey & Company, Fourth Industrial Revolution: Beacons of Technology and Innovation in Manufacturing, white paper, January 2019, p. 19, retrieved 2026-11-02: https://www3.weforum.org/docs/WEF_4IR_Beacons_of_Technology_and_Innovation_in_Manufacturing_report_2019.pdf
• William Samuelson and Richard Zeckhauser, "Status Quo Bias in Decision Making," Journal of Risk and Uncertainty 1(1), 7–59, 1988, abstract retrieved 2026-11-02: https://ideas.repec.org/a/kap/jrisku/v1y1988i1p7-59.html

On the Zapier survey. Centiment fielded it for Zapier from May 8 to 15, 2026, with 835 completed responses from U.S. professionals at manager level or above who had direct knowledge of their company's AI pilots, at companies with 100 or more employees that were testing, piloting, or using AI in at least one workflow. Data is unweighted. The stall reasons are reported here as Zapier states them; Zapier gives some as shares of stalled pilots and others as shares of organizations, and respondents could give more than one. Zapier sells automation software.

On the World Economic Forum figures. The 2019 white paper covers Fourth Industrial Revolution technology in manufacturing broadly (including AI, analytics, connected sensors, and 3D printing), not AI alone. The "more than 70%," 29%, and pilot-length figures (15% under one year, 56% one to two years, 28% over two years) are on p. 19 and are attributed to the Forum's earlier survey work.

On the electrical contractor. It's an illustrative composite, not a client. Its figures (85 people, three pilots, 14, 10, and 7 months, about 20 chats a month, $1,150 a month, roughly $12,000, nine and three business days, 15 and ten minutes of editing, ten and four booked calls, 50 submittals, 17 mismatches, 11 caught, 60% scanned, six weeks) are representative examples, not measurements or benchmarks.

---

About Avolis Research Group

Avolis Research Group is Avolis's in-house research practice, focused on how operations-heavy small and mid-sized businesses actually adopt AI. It synthesizes primary economic research, government survey data, and results from real implementations into practical, vendor-neutral guidance.

More about Avolis and how we work · Get in touch